Built here · UAE
MBZUAI's IFM and Inception42 launch Jais 3, six open Arabic models and a speech model
The developers say Jais 3 includes six language models, from 900 million to 375 billion parameters, plus Jais Speech. Their performance claims come from their own evaluations.
What was launched
Middle East AI News reported on 15 October 2026 that MBZUAI's Institute of Foundation Models (IFM) and Inception42 launched Jais 3. It is a fleet of six open Arabic language models plus Jais Speech, a speech recognition model. Both the language models and the speech model are delivered, not only announced.1,2
MBZUAI is the Mohamed bin Zayed University of Artificial Intelligence. Inception42 is G42's sovereign agentic AI company. Middle East AI News also reports that the family was developed in collaboration with Cerebras Systems.1,2
The six language models range from 900 million to 375 billion parameters. Parameters are the internal settings a model learns in training. For scale, the Stanford AI Index lists PaLM, from 2022, at 540 billion. The models are built on K2 Horizon, IFM's open-source foundation model fleet released in September 2026.1,2,3
Dialects and speech
The models are trained on Modern Standard Arabic and regional dialects, from Emirati and Gulf to North African Arabic. Jais Speech covers dialects from 19 Arab countries.1,2
When speakers switch languages, Jais Speech keeps English or French words in their original script. It also identifies the dialect and can add Arabic vowel marks. IFM's model card also lists natural switching into Spanish and Italian.1,2,4
Performance claims
The developers say all models share one training recipe, agent capabilities and a 512,000-token context window, which they describe as the longest of any Arabic-native model. They say they evaluated the models on an audited suite of 24 Arabic benchmarks, which they are releasing publicly.1,2
According to the developers, the 375-billion-parameter model leads open models on agentic tasks and matches the strongest on Arabic understanding. They say the 3.7-billion-parameter model outperformed every open model of up to 5 billion parameters. They also say the 7B and 32B models had the highest average Arabic scores among open models at their sizes.1,2
At 7 billion parameters, the developers report their average Arabic score rose from 56.4 for the K2 Horizon base model to 79.6 for Jais 3. They also say Jais Speech recorded the lowest average word and character error rates of any model compared, including commercial services.1,2
Access and licenses
Each model ships with five training checkpoints for researchers, and the models are available on Hugging Face. Middle East AI News links jaischat.ai as the chat demo.1,2
The model cards list the 375-billion-parameter model and the speech model weights under the open Apache 2.0 license.4,5
What could not be verified
- All performance and ranking claims come from the developers' own evaluations, with no independent verification.
- The average Arabic score rising from 56.4 to 79.6 at 7 billion parameters is the developers' own figure.
- The claim of the longest context window of any Arabic-native model, 512,000 tokens, is the developers' own.
- Whether the 24 benchmarks have been published or only promised is not clear.
- Real-world accuracy of Jais Speech across the dialects of 19 countries has not been independently tested.
Sources